Location-aware convolutional neural networks for graph classification.

Area of Science:

• Graph Neural Networks
• Machine Learning
• Computer Science

Background:

• Graph classification tasks rely heavily on identifying graph patterns, such as chemical patterns in molecular graphs.
• Convolutional Neural Networks (CNNs) are adapted for graph classification due to their pattern learning capabilities.
• Challenges in applying CNNs to graphs include varying node neighborhoods and lack of canonical node ordering, leading to information loss in existing heuristic methods.

Purpose of the Study:

• To propose a novel approach for graph classification that overcomes limitations of existing methods.
• To develop a Location learning-based Convolutional Neural Network (LCNN) that retains all node information and learns task-oriented patterns.
• To enhance the effectiveness of CNNs in graph classification by addressing challenges in receptive field construction.

Main Methods:

• Proposed Location learning-based Convolutional Neural Networks (LCNN) for graph classification.
• LCNN constructs receptive fields by learning node locations based on embeddings containing structural and feature information.
• Standard CNNs are then applied to capture graph patterns using the learned node locations.

Main Results:

• Experimental results demonstrate the effectiveness of the proposed LCNN method.
• The location learning mechanism retains information from all nodes, unlike heuristic methods that drop nodes.
• LCNN exhibits a strong ability for task-oriented pattern learning in graph classification.

Conclusions:

• LCNN offers an effective solution for graph classification by learning node locations and retaining complete graph information.
• The proposed method enhances CNNs' capability for graph pattern learning in a task-oriented manner.
• LCNN represents a significant advancement in applying deep learning to graph-based classification problems.